Effects of case removal in prognostic models

Area of Science:

• Machine Learning in Predictive Modeling
• Statistical Learning and Data Mining

Background:

• Building and updating prognostic models requires extensive training data, which is time-consuming.
• Compact and informative training sets accelerate model development and evaluation.
• Efficient methods for selecting and removing training cases are crucial for prognostic model construction.

Purpose of the Study:

• To compare different regression diagnostic methods for selecting and removing training cases in prognostic models.
• To evaluate the effectiveness of univariate, sequential multivariate, and genetic algorithm approaches for case selection.

Main Methods:

• Univariate case selection using classical regression diagnostic statistics.
• Multivariate case selection using a sequential backward elimination approach.
• Multivariate case selection using a non-sequential genetic algorithm.

Main Results:

• The genetic algorithm approach resulted in final models with fewer cases while maintaining predictive capability.
• Univariate and sequential multivariate methods were less effective in identifying optimal case subsets.
• Genetic algorithms demonstrated an ability to detect influential case sets that might be missed by isolated analysis.

Conclusions:

• Genetic algorithms offer a superior approach for case selection in training sets for prognostic models compared to univariate or sequential multivariate methods.
• This method aids in creating more compact yet informative training sets, thereby streamlining model building and evaluation.
• The effectiveness of genetic algorithms lies in their capacity to identify groups of cases influential collectively.